Speaker
Description
The Virtual Solar Observatory (VSO) indexes instruments spanning X-ray through radio wavelengths, yet discovery remains largely keyword-dependent — requiring users to know provider names and instrument codes in advance. We present a prototype semantic search layer built on top of VSO that replaces exact-match lookup with embedding-based similarity search, enabling researchers to find relevant instruments through natural language queries and conceptual proximity rather than precise terminology.
Each instrument record was encoded into a high-dimensional vector using a pre-trained large language model, with cosine similarity used to rank the most semantically related instruments for any given query. A Principal Component Analysis projection reveals natural clustering by wavelength band — with EUV, UV, X-ray, and radio instruments forming visually distinct groupings — confirming that the embedding space captures physically meaningful structure without VSO-specific supervision.
Early results demonstrate that the model surfaces cross-mission instrument equivalents that keyword search would miss entirely, such as associating SDO/AIA EUV channels with complementary observations from STEREO/EUVI and Proba-2/SWAP. This work lays the groundwork for a conversational data discovery interface where heliophysicists can describe the data they need in scientific terms and receive ranked, cross-archive instrument recommendations — complementing the structured API and TAP access capabilities being developed in VSO 2.0.